Enhanced Acoustic Echo Suppression Techniques Based on Spectro-Temporal Correlations

Abstract

In the past decades, a number of approaches have been dedicated to acoustic echo cancellation and suppression which reduce the negative effects of acoustic echo, namely the acoustic coupling between the loudspeaker and microphone in a room. In particular, the increasing use of full-duplex telecommunication systems has led to the requirement of faster and more reliable acoustic echo cancellation algorithms. The solutions have been based on adaptive filters, but the length of these filters has to be long enough to consider most of the echo signal and linear filtering in these algorithms may be limited to remove the echo signal in various environments.

In this thesis, a novel stereophonic acoustic echo suppression (SAES) technique based on spectral and temporal correlations is proposed in the short-time Fourier transform (STFT) domain. Unlike traditional stereophonic acoustic echo cancellation, the proposed algorithm estimates the echo spectra in the STFT domain and uses a Wiener filter to suppress echo without performing any explicit double-talk detection. The proposed approach takes account of interdependencies among components in adjacent time frames and frequency bins, which enables more accurate estimation of the echo signals.

Due to the limitations of power amplifiers or loudspeakers, the echo signals captured in the microphones are not in a linear relationship with the far-end signals even when the echo path is perfectly linear. The nonlinear components of the echo cannot be successfully removed by a linear acoustic echo canceller. The remaining echo components in the output of acoustic echo suppression (AES) can be further suppressed by applying residual echo suppression (RES) algorithms. In this thesis, we propose an optimal RES gain estimation based on deep neural network (DNN) exploiting both the far-end and the AES output signals in all frequency bins. A DNN structure is introduced as a regression function representing the complex nonlinear mapping from these signals to the optimal RES gain. Because of the capability of the DNN, the spectro-temporal correlations in the full-band can be considered while finding the nonlinear function. The proposed method does not require any explicit double-talk detectors to deal with single-talk and double-talk situations.

One of the well-known approaches for nonlinear acoustic echo cancellation is an adaptive Volterra filtering and various algorithms based on the Volterra filter were proposed to describe the characteristics of nonlinear echo and showed the better performance than the conventional linear filtering. However, the performance might be not satisfied since these algorithms could not consider the full correlation for the nonlinear relationship between the input signal and far-end signal in time-frequency domain. In this thesis, we propose a novel DNN-based approach for nonlinear acoustic echo suppression (NAES), extending the proposed RES algorithm. Instead of estimating the residual gain for suppressing the nonlinear echo components, the proposed algorithm straightforwardly recovers the near-end speech signal through the direct gain estimation obtained from DNN frameworks on the input and far-end signal. For echo aware training, a priori and a posteriori signal-to-echo ratio (SER) are introduced as additional inputs of the DNN for tracking the change of the echo signal. In addition, the multi-task learning (MTL) to the DNN-based NAES is combined to the DNN incorporating echo aware training for robustness. In the proposed system, an additional task of double-talk detection is jointly trained with the primary task of the gain estimation for NAES. The DNN can learn the good representations which can suppress more in single-talk periods and improve the gain estimates in double-talk periods through the MTL framework. Besides, the proposed NAES using echo aware training and MTL with double-talk detection makes the DNN be more robust in various conditions.

The proposed techniques show significantly better performance than the conventional AES methods in both single- and double-talk periods. As a pre-processing of various applications such as speech recognition and speech enhancement, these approaches can help to transmit the clean speech and provide an acceptable communication in full-duplex real environments.